Advanced Synthetic Data Generation for ADAS: Infrastructure Sensors and CARLA-Omniverse Integration

Advanced Driver Assistance Systems (ADAS) validation necessitates large datasets from a variety of driving situations, yet conventional data gathering techniques have serious drawbacks in terms of cost, safety, and scenario coverage. This thesis builds on existing CARLA-based frameworks for generating synthetic data by adding two important new features aimed at improving dataset quality and realism. First, an infrastructure-based sensor simulation system was created, expanding beyond vehicle-mounted sensors to include fixed sensors on traffic lights and road infrastructure, with a dedicated management tool for dynamic sensor placement and configuration while retaining nuScenes dataset compatibility. Second, a novel framework was developed to export complete CARLA simulations to Universal Scene Description (USD) format for integration with NVIDIA Isaac Sim. This framework captures entire simulations as temporal sequences with keyframed animations, allowing deterministic replay in Omniverse's photorealistic rendering environment. An Isaac Sim extension was created to facilitate the management of imported simulations and the collection of sensor data in nuScenes format. CARLA's official USD export framework for version 0.9.16 was announced during development, prompting the strategic decision to prepare for integration with the forthcoming official solution. Despite the fact that the custom USD implementation has material limitations in the current visual output, the framework that has been developed establishes a robust integration pipeline between CARLA and Isaac Sim. This pipeline will facilitate improved photorealistic rendering capabilities and enhanced synchronization when combined with the official CARLA USD framework. This methodology offers a scalable foundation for ADAS validation that maintains industry-standard compatibility and combines the potential for near-photorealistic sensor simulation with the flexibility of virtual environments.